Yiwei Wang

DICE transforms long-document retrieval by effectively preserving critical information from chunks, achieving up to a 60% increase in retrieval accuracy for documents over 4,000 tokens.

SelectStream reveals that selective memory allocation can dramatically enhance streaming video understanding, outperforming traditional methods by preserving scene perception without sacrificing performance.

Rationale-based fine-tuning may actually undermine clinical prediction accuracy, challenging the belief that teaching models "why" can enhance their performance.

LLMs can now reliably generate complex, editable SVG diagrams by explicitly optimizing for geometric constraints via reinforcement learning, opening the door to automated technical illustration.

TwinGate stops jailbreaks by tracking malicious intent across anonymized, interleaved queries with minimal overhead, something previous defenses couldn't do.

Energy-dissipation principles can revolutionize how we infer potential functions in noisy, incomplete data environments, achieving remarkable robustness in generalized diffusion processes.

LLMs get *more* creative at generating molecules when you add *more* constraints, defying the intuition that creativity thrives on freedom.

Forget flattening: VideoStir's spatio-temporal graph retrieval and intent-aware scoring unlocks more effective reasoning over long videos.

Prompt highlighting in LLMs gets a serious upgrade: PRISM-$\Delta$ steers models to focus on relevant text spans with better accuracy and fluency, even in long contexts.

Forget expensive retraining: PromptCD unlocks significant improvements in LLM alignment and VLM visual grounding simply by contrasting model responses to cleverly designed prompts at test time.